SPAC: Sampling-Based Progressive Attribute Compression for Dense Point Clouds.

Area of Science:

• Computer Vision
• 3D Data Processing
• Signal Processing

Background:

• Dense point clouds are crucial for 3D applications but require efficient compression.
• Existing geometry-based point cloud compression (G-PCC) standards face challenges in attribute compression efficiency.
• Learning-based methods offer potential for improved point cloud data compression.

Purpose of the Study:

• To develop an end-to-end attribute compression method for dense point clouds.
• To enhance compression efficiency and reconstruction accuracy compared to existing standards.
• To introduce a novel learning-based approach that outperforms the G-PCC standard.

Main Methods:

• Frequency sampling using Fast Fourier Transform (FFT) and Hamming window.
• Octree partitioning for structured sub-point cloud processing.
• Adaptive scale feature extraction with geometry assistance and offset-attention.
• Global hyperprior entropy model for efficient encoding.
• Mirrored network decoder for progressive feature restoration.

Main Results:

• Achieved an average Bjøntegaard delta bitrate reduction of -24.58% (Y component) and -21.23% (YUV components) on the MPEG Category Solid dataset.
• Achieved an average Bjøntegaard delta bitrate reduction of -22.48% (Y component) and -17.19% (YUV components) on the MPEG Category Dense dataset.
• Demonstrated superior performance over the latest G-PCC standard on dense point cloud datasets.

Conclusions:

• The proposed learning-based attribute codec is the first to outperform the G-PCC standard under common test conditions.
• The method offers significant improvements in compression efficiency and reconstruction quality for dense point clouds.
• The developed technique paves the way for more efficient 3D data transmission and storage.